| Message ID | 1425333797-26082-1-git-send-email-tariq.x.saeed@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
> Ocfs2-devel] [PATCH 1/1] BUG_ON(lockres->l_level != DLM_LOCK_EX && > !checkpointed) tripped in ocfs2_ci_checkpointed As I was working on this patch, I noticed that the changes in ocfs2_setattr are done in more than one transaction --> in case of crash somewhere in this func, say after doing the first trans, the inode could be in a state where only some of the changes have been applied (after log replay). Checking ext3_ocfs2, from where this code was ported, I found the same issue there as well. Do you have any comments? Thanks -Tariq Saeed PS I realize that doing everything in ocfs2_setattr in a single transaction is a major surgery as the functions called down the path for xattr start their own transactions with different credits.
Hi Tariq, Thanks for the patch, btw if you want your patches to go upstream you'll have a much better chance if you CC Me and Andrew (along with the mailing list). Comments inline: On Mon, Mar 02, 2015 at 02:03:17PM -0800, Tariq Saeed wrote: > Orabug: 20189959 > > PID: 614 TASK: ffff882a739da580 CPU: 3 COMMAND: "ocfs2dc" > #0 [ffff882ecc3759b0] machine_kexec at ffffffff8103b35d > #1 [ffff882ecc375a20] crash_kexec at ffffffff810b95b5 > #2 [ffff882ecc375af0] oops_end at ffffffff815091d8 > #3 [ffff882ecc375b20] die at ffffffff8101868b > #4 [ffff882ecc375b50] do_trap at ffffffff81508bb0 > #5 [ffff882ecc375ba0] do_invalid_op at ffffffff810165e5 > #6 [ffff882ecc375c40] invalid_op at ffffffff815116fb > [exception RIP: ocfs2_ci_checkpointed+208] > RIP: ffffffffa0a7e940 RSP: ffff882ecc375cf0 RFLAGS: 00010002 > RAX: 0000000000000001 RBX: 000000000000654b RCX: ffff8812dc83f1f8 > RDX: 00000000000017d9 RSI: ffff8812dc83f1f8 RDI: ffffffffa0b2c318 > RBP: ffff882ecc375d20 R8: ffff882ef6ecfa60 R9: ffff88301f272200 > R10: 0000000000000000 R11: 0000000000000000 R12: ffffffffffffffff > R13: ffff8812dc83f4f0 R14: 0000000000000000 R15: ffff8812dc83f1f8 > ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018 > #7 [ffff882ecc375d28] ocfs2_check_meta_downconvert at ffffffffa0a7edbd [ocfs2] > #8 [ffff882ecc375d38] ocfs2_unblock_lock at ffffffffa0a84af8 [ocfs2] > #9 [ffff882ecc375dc8] ocfs2_process_blocked_lock at ffffffffa0a85285 [ocfs2] > #10 [ffff882ecc375e18] ocfs2_downconvert_thread_do_work at ffffffffa0a85445 [ocfs2] > #11 [ffff882ecc375e68] ocfs2_downconvert_thread at ffffffffa0a854de [ocfs2] > #12 [ffff882ecc375ee8] kthread at ffffffff81090da7 > #13 [ffff882ecc375f48] kernel_thread_helper at ffffffff81511884 > > assert is tripped because the tran is not checkpointed and the lock level is PR. > Following call could leave the lock in PR state, latter on causing the assert. > ocfs2_setattr -> ocfs2_acl_chmod -> ocfs2_set_acl -> ocfs2_acl_set_mode > Here is how. > > 1116 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr) > 1117 { > 1261 ocfs2_inode_unlock(inode, 1); <<< > .. > 1272 if (!status && attr->ia_valid & ATTR_MODE) { > 1273 status = ocfs2_acl_chmod(inode); You're looking at an old kernel, that's now a call to posix_acl_chmod(). You'll want to send patches against a recent kernnel too. > 330 int ocfs2_acl_chmod(struct inode *inode) > 351 ret = ocfs2_set_acl(NULL, inode, NULL, ACL_TYPE_ACCESS, > > 230 static int ocfs2_set_acl(handle_t *handle, > 231 struct inode *inode, > 232 struct buffer_head *di_bh, > 233 int type, > 258 ret = ocfs2_acl_set_mode(inode, di_bh, > > 175 static int ocfs2_acl_set_mode(struct inode *inode, struct buffer_head *di_bh, > 176 handle_t *handle, umode_t new_mode) > 177 { > 178 int ret, commit_handle = 0; > 179 struct ocfs2_dinode *di; > 180 > 181 if (di_bh == NULL) { > 182 ret = ocfs2_read_inode_block(inode, &di_bh); > 183 if (ret) { > 184 mlog_errno(ret); > 185 goto out; > 186 } > 187 } else > 188 get_bh(di_bh); > 189 > 190 if (handle == NULL) { > >> BUG: inode lock not held in ex at this point << > 191 handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb), > > ocfs2_setattr.#1261 we unlock and at #1273 call ocfs2_acl_chmod. When we reach > ocfs2_acl_set_mode.#191 and do trans, the inode cluster lock is not held in EX > mode (it should be). How this could have happended? Ok, so the problem is that we're not holding a cluster lock across acl operations. This actually affects calls from the VFS as well. You have the right idea here - make sure we do this under lock :) We want to do the locking in ocfs2_iop_set_acl() and ocfs2_iop_get_acl() though. Those both get called from the VFS and the path from posix_acl_chmod() will hit that too, so we can kill two birds with one stone. Thanks, --Mark -- Mark Fasheh
On Mon, Mar 02, 2015 at 05:30:51PM -0800, Tariq Saeed wrote: >> Ocfs2-devel] [PATCH 1/1] BUG_ON(lockres->l_level != DLM_LOCK_EX && >> !checkpointed) tripped in ocfs2_ci_checkpointed > As I was working on this patch, I noticed that the changes in > ocfs2_setattr are done in more than > one transaction --> in case of crash somewhere in this func, say after > doing the first trans, > the inode could be in a state where only some of the changes have > been applied (after log replay). > > Checking ext3_ocfs2, from where this code was ported, I found the same > issue there as well. Can you give a specific example where this would cause a problem for the user? > Do you have any comments? Well other than my question above, some things are impossible to always fit in one transaction. Extends and truncates are a good example of this. Ext4 does orphan inodes during truncate so it can complete them on remount in case of crash. Ocfs2 doesn't do this though it could be done pretty easily so if you're interested, that'd be a nice feature for us to finally cross off the list :) --Mark -- Mark Fasheh
On 03/04/2015 03:57 PM, Mark Fasheh wrote:
> Ok, so the problem is that we're not holding a cluster lock across acl
Yes, but in addition to that, the inode lock in ocfs2_setattr MUST not
be dropped till
the exit since all operations,mode change, size change etc,
in that function plus any changes to acls are atomic from user's point
of view.
Thanks,
-Tariq Saeed
On Thu, Mar 05, 2015 at 11:33:55AM -0800, Tariq Saeed wrote: > > On 03/04/2015 03:57 PM, Mark Fasheh wrote: > > Ok, so the problem is that we're not holding a cluster lock across acl > Yes, but in addition to that, the inode lock in ocfs2_setattr MUST not > be dropped till > the exit since all operations,mode change, size change etc, > in that function plus any changes to acls are atomic from user's point > of view. That's fine, my primary concern is that we're not locking ACL calls _at all_. The fix you submitted doesn't really adress that, it just fixes one isolated case, hence my description of a more complete solution :) --Mark -- Mark Fasheh
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c index 3950693..113880c 100644 --- a/fs/ocfs2/file.c +++ b/fs/ocfs2/file.c @@ -1118,7 +1118,7 @@ out: int ocfs2_setattr(struct dentry *dentry, struct iattr *attr) { - int status = 0, size_change; + int status = 0, size_change, inode_locked = 0; struct inode *inode = dentry->d_inode; struct super_block *sb = inode->i_sb; struct ocfs2_super *osb = OCFS2_SB(sb); @@ -1164,6 +1164,7 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr *attr) mlog_errno(status); goto bail_unlock_rw; } + inode_locked = 1; if (size_change) { status = inode_newsize_ok(inode, attr->ia_size); @@ -1244,7 +1245,10 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr *attr) bail_commit: ocfs2_commit_trans(osb, handle); bail_unlock: - ocfs2_inode_unlock(inode, 1); + if (status) { + ocfs2_inode_unlock(inode, 1); + inode_locked = 0; + } bail_unlock_rw: if (size_change) ocfs2_rw_unlock(inode, 1); @@ -1260,6 +1264,8 @@ bail: if (status < 0) mlog_errno(status); } + if (inode_locked) + ocfs2_inode_unlock(inode, 1); return status; }
Orabug: 20189959 PID: 614 TASK: ffff882a739da580 CPU: 3 COMMAND: "ocfs2dc" #0 [ffff882ecc3759b0] machine_kexec at ffffffff8103b35d #1 [ffff882ecc375a20] crash_kexec at ffffffff810b95b5 #2 [ffff882ecc375af0] oops_end at ffffffff815091d8 #3 [ffff882ecc375b20] die at ffffffff8101868b #4 [ffff882ecc375b50] do_trap at ffffffff81508bb0 #5 [ffff882ecc375ba0] do_invalid_op at ffffffff810165e5 #6 [ffff882ecc375c40] invalid_op at ffffffff815116fb [exception RIP: ocfs2_ci_checkpointed+208] RIP: ffffffffa0a7e940 RSP: ffff882ecc375cf0 RFLAGS: 00010002 RAX: 0000000000000001 RBX: 000000000000654b RCX: ffff8812dc83f1f8 RDX: 00000000000017d9 RSI: ffff8812dc83f1f8 RDI: ffffffffa0b2c318 RBP: ffff882ecc375d20 R8: ffff882ef6ecfa60 R9: ffff88301f272200 R10: 0000000000000000 R11: 0000000000000000 R12: ffffffffffffffff R13: ffff8812dc83f4f0 R14: 0000000000000000 R15: ffff8812dc83f1f8 ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018 #7 [ffff882ecc375d28] ocfs2_check_meta_downconvert at ffffffffa0a7edbd [ocfs2] #8 [ffff882ecc375d38] ocfs2_unblock_lock at ffffffffa0a84af8 [ocfs2] #9 [ffff882ecc375dc8] ocfs2_process_blocked_lock at ffffffffa0a85285 [ocfs2] #10 [ffff882ecc375e18] ocfs2_downconvert_thread_do_work at ffffffffa0a85445 [ocfs2] #11 [ffff882ecc375e68] ocfs2_downconvert_thread at ffffffffa0a854de [ocfs2] #12 [ffff882ecc375ee8] kthread at ffffffff81090da7 #13 [ffff882ecc375f48] kernel_thread_helper at ffffffff81511884 assert is tripped because the tran is not checkpointed and the lock level is PR. Following call could leave the lock in PR state, latter on causing the assert. ocfs2_setattr -> ocfs2_acl_chmod -> ocfs2_set_acl -> ocfs2_acl_set_mode Here is how. 1116 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr) 1117 { 1261 ocfs2_inode_unlock(inode, 1); <<< .. 1272 if (!status && attr->ia_valid & ATTR_MODE) { 1273 status = ocfs2_acl_chmod(inode); 330 int ocfs2_acl_chmod(struct inode *inode) 351 ret = ocfs2_set_acl(NULL, inode, NULL, ACL_TYPE_ACCESS, 230 static int ocfs2_set_acl(handle_t *handle, 231 struct inode *inode, 232 struct buffer_head *di_bh, 233 int type, 258 ret = ocfs2_acl_set_mode(inode, di_bh, 175 static int ocfs2_acl_set_mode(struct inode *inode, struct buffer_head *di_bh, 176 handle_t *handle, umode_t new_mode) 177 { 178 int ret, commit_handle = 0; 179 struct ocfs2_dinode *di; 180 181 if (di_bh == NULL) { 182 ret = ocfs2_read_inode_block(inode, &di_bh); 183 if (ret) { 184 mlog_errno(ret); 185 goto out; 186 } 187 } else 188 get_bh(di_bh); 189 190 if (handle == NULL) { >> BUG: inode lock not held in ex at this point << 191 handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb), ocfs2_setattr.#1261 we unlock and at #1273 call ocfs2_acl_chmod. When we reach ocfs2_acl_set_mode.#191 and do trans, the inode cluster lock is not held in EX mode (it should be). How this could have happended? We are the lock master, were holding lock EX and have released it in ocfs2_setattr.#1261. Note that there are no holders of this lock at this point. Another node needs the lock in PR, and we downconvert from EX to PR. So the inode lock is PR when do the trans in ocfs2_acl_set_mode.#191. The trans stays in core (not flushed to disc). Now another node want the lock in EX, downconvert thread gets kicked (the one that tripped assert abovt), finds an unflushed trans but the lock is not EX (it is PR). If the lock was at EX, it would have flushed the trans ocfs2_ci_checkpointed -> ocfs2_start_checkpoint before downconverting (to NULL) for the request. ocfs2_setattr must retain inode lock ex in this code path from. Signed-off-by: Tariq Saeed <tariq.x.saeed@oracle.com> --- fs/ocfs2/file.c | 10 ++++++++-- 1 files changed, 8 insertions(+), 2 deletions(-)